MISTRAL: a lattice translation system for IWSLT 2007

MISTRAL: A Lattice Translation System for IWSLT 2007 Alexandre Patry, Philippe Langlais RALI – DIRO Universit´e de Montr´eal {patryale,felipe}@iro.umontreal.ca Fr´ed´eric B´echet LIA University of Avignon frederic.bechet@univ-avignon.fr Abstract This paper describes MISTRAL, the lattice translation system that we developed for the Italian-English track of the Inter- national Workshop on Spoken Language Translation 2007. MISTRAL is a discriminative phrase-based system that translates a source word lattice in two passes. The first pass extracts a list of top ranked sentence pairs from the lattice and the second pass rescores this list with more complex fea- tures. Our experiments show that our system, when trans- lating pruned lattices, is at least as good as a fair baseline that translates the first ranked sentences returned by a speech recognition system. 1. Introduction While in textual translation the sentences to translate are known for sure, in spoken language translation they must be reconstructed from an audio signal. Instead of translating the raw audio signal, spoken language translation systems usu- ally start from the output of an automatic speech recognition system, which usually takes the form of a word lattice. A straightforward way to translate a word lattice is to feed a textual translation system with the most probable sen- tence. To overcome some recognition errors, one can in- stead translate a set of n top ranked sentences extracted from the lattice and select the best translation among the one re- turned [1, 2]. Finally, a third approach is to translate directly the lattice with a specialized decoder tightly coupled with the speech recognition system [3, 4, 5, 6, 7]. In this paper, we present MISTRAL (Monotone yet Im- perfect Statistical TRAnslation of Lattices), a discriminative phrase-based system that translates lattices in two passes. The first pass uses a beam-search decoder to extract a N-Best list of source sentences and their translations from the lat- tice and the second pass rescores this list with more complex feature functions. Different word-based lattice translation systems are pre- sented in [3, 4, 5]. A generative phrase-based system is de- scribed in [6]. This system translates the lattice with a se- quence of weighted finite state machines applied in cascade while our system is dedicated and traverses the lattice only once. In [7], the authors present a discriminative phrase- based system that translates confusion networks and offers an elegant solution to word reordering. The remaining of the article is organized as follow. The next section presents the theoretical framework of spoken language translation. We describe our decoder in section 3 and evaluate our complete system in section 4. We finally conclude in section 5. 2. Lattice Translation In textual translation, a source sentence (f) is known for a fact, we thus seek to resolve: e⋆ = argmax e Pr(e|f) (1) Since in spoken language translation only a word lattice is known (o), the equation to resolve thus becomes: e⋆ = argmax e � f Pr(e, f|o) (2) If we pose the assumption that the target sentence and the word lattice are …

Content truncated to 500 words